Is open-source AI safer or riskier than closed/proprietary AI?

Open vs closed AI safety is hotly debated — I'll weigh the evidence for each model's risks and benefits.

Open-source AI is not inherently safer or riskier than closed / proprietary AI - it is safer in some ways and riskier in others. The consensus in recent analysis is that neither is automatically safer

> Overall, open source software is neither inherently more nor less secure than proprietary software.

> Is open‑source AI safer than closed‑source AI? ... Not inherently — it’s safer in some ways and riskier in others.

Which is safer depends on deployment design, governance, data flow, access control, vendor terms, and internal operational maturity, and that:

> The safer option is not defined by whether the model is open or closed. It is defined by how well the model fits your data sensitivity, compliance needs, governance maturity, and risk controls.

Why open-source is argued to be SAFER

1. Transparency and auditability

> Researchers to audit AI for biases, misinformation risks, and ethical concerns... Developers to modify models for greater fairness and safety. Regulators to evaluate AI impacts without relying on corporate transparency.

> Transparency and auditability: Openweight AI models enable independent scrutiny of model ... This transparency supports oversight, compliance assessments, and research on robustness, security, and safety
and
> transparency and auditability associated with open models can strengthen oversight, compliance, and trust.

2. Security through community, not obscurity

Open-source advocates argue that "security through obscurity" is a fallacy

> They contend that by keeping models closed, the industry is ignoring systemic risks that only a global community of researchers can identify and fix. ... From this perspective, open-source AI is actually safer because it eliminates the single point of failure and prevents a small group of engineers from having unilateral control

> Their scrutiny by the broader community facilitates the early detection and patching of vulnerabilities, making these models potentially safer than their closed counterparts

This argument was echoed by Mark Zuckerberg in 2024, who expected open source AI to be safer than alternatives, especially against unintentional harm, because open systems can be widely scrutinized.

3. Counter to concentration of power

Proprietary services such as ChatGPT, Gemini, and Claude are not inherently safer: their operators can keep flaws from public scrutiny, monitor how people use them, impose restrictions, raise prices, or withdraw access. Open models reduce vendor lock-in and allow private deployment in VPC, on-premises or edge, giving organizations higher control over data residency, logging and auditability when deployed privately.

Evidence presented at the International Conference on Machine Learning found that available evidence does not show that open models are riskier than closed models, partly because even closed models can be subverted and information for malicious actors might already be available on the internet.

Why open-source is argued to be RISKIER

1. Irreversible proliferation and loss of central control

> They are also just a file – very expensive to produce, but trivial to copy, and impossible to recall once distributed.
> Major frontier AI companies such as OpenAI, Anthropic and Google do not release the weights of their frontier models. This means they alone can patch the model, add or tighten safeguards, and as a last resort remove the model from circulation entirely, as happened in June 2026 when the US government directed Anthropic to suspend access to Fable 5. An open-weight model, on the other hand, has no such restrictions – the developer publishes the weights and thereafter has no meaningful ability to recall, patch or supervise any copies.

2. Easy removal of safety guardrails

In-model safety training (refusals learned through alignment) travels with the weights but is fragile, and external safety classifiers that screen prompts are not compulsory for open-weight deployments:

> There is nothing to compel an open-weight provider to run them, and given that there are many providers that supply access to the same models, a determined malicious actor can search for a deployment that does not run with those safeguards.

Researchers have demonstrated how cheap stripping is:

> Researchers at Palisade Research removed the safety training from Meta’s Llama 3 8B in around five minutes on a single GPU for under a dollar, and from the 70B model in well under an hour

More broadly:

> It is now relatively inexpensive to fine-tune an open-weight model to remove the in-model safety layer, meaning such models are much easier to use for criminal activities.

3. Lower barrier to misuse for high-impact harms

The UK’s National Cyber Security Centre assessed that AI will ‘almost certainly’ continue to make cyber intrusion more effective and efficient, and that the proliferation of open source and commercially available models will lower the barrier to entry for a widening range of actors.

Compared to closed-weight models:

> In contrast, open-weight models introduce new and unique risks – they provide near-frontier capabilities which could be used for criminal purposes, are widely available to anyone who intends harm, and are easier to misuse

Concerns most often cited are:

4. Transparency does not solve all harms

> Technically speaking, while open-source AI – however we define it – may provide a higher degree of transparency compared to proprietary AI technologies, other risks including hallucinations and inaccuracies, capacity to cause harm, infringement of intellectual property, and privacy risks are not necessarily mitigated in any manner by the fact that AI is an open source AI

Why closed / proprietary is argued to be SAFER (and its limits)

Closed models allow:

However, closed is not a guarantee of safety:

> Even closed models can be subverted and... information for malicious actors might already be available on the internet through search engines.

Recent incident reports from Anthropic and OpenAI in 2026 also showed labs running frontier models internally without the safety monitoring used on public versions, raising concerns about internal safety culture behind closed doors. Closed systems also create single points of failure, concentration of power, and auditability problems for regulators and independent researchers.

Current policy consensus

Most regulators and recent G7/OECD, EU AI Act, and UK RUSI analyses now treat the question as a trade-off to be managed rather than a binary choice:

For many enterprises this leads to a hybrid strategy: using proprietary APIs for frontier reasoning with vendor-managed controls, and open-weight models for private, high-volume, or regulated workloads where data control and auditability matter most.